86 lines
2.8 KiB
Python
86 lines
2.8 KiB
Python
# -*- coding: utf-8 -*-
|
|
"""
|
|
sleekxmpp.xmlstream.matcher.xpath
|
|
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
|
|
|
Part of SleekXMPP: The Sleek XMPP Library
|
|
|
|
:copyright: (c) 2011 Nathanael C. Fritz
|
|
:license: MIT, see LICENSE for more details
|
|
"""
|
|
|
|
from sleekxmpp.xmlstream.stanzabase import ET
|
|
from sleekxmpp.xmlstream.matcher.base import MatcherBase
|
|
|
|
|
|
# Flag indicating if the builtin XPath matcher should be used, which
|
|
# uses namespaces, or a custom matcher that ignores namespaces.
|
|
# Changing this will affect ALL XPath matchers.
|
|
IGNORE_NS = False
|
|
|
|
|
|
class MatchXPath(MatcherBase):
|
|
|
|
"""
|
|
The XPath matcher selects stanzas whose XML contents matches a given
|
|
XPath expression.
|
|
|
|
.. warning::
|
|
|
|
Using this matcher may not produce expected behavior when using
|
|
attribute selectors. For Python 2.6 and 3.1, the ElementTree
|
|
:meth:`~xml.etree.ElementTree.Element.find()` method does
|
|
not support the use of attribute selectors. If you need to
|
|
support Python 2.6 or 3.1, it might be more useful to use a
|
|
:class:`~sleekxmpp.xmlstream.matcher.stanzapath.StanzaPath` matcher.
|
|
|
|
If the value of :data:`IGNORE_NS` is set to ``True``, then XPath
|
|
expressions will be matched without using namespaces.
|
|
"""
|
|
|
|
def match(self, xml):
|
|
"""
|
|
Compare a stanza's XML contents to an XPath expression.
|
|
|
|
If the value of :data:`IGNORE_NS` is set to ``True``, then XPath
|
|
expressions will be matched without using namespaces.
|
|
|
|
.. warning::
|
|
|
|
In Python 2.6 and 3.1 the ElementTree
|
|
:meth:`~xml.etree.ElementTree.Element.find()` method does not
|
|
support attribute selectors in the XPath expression.
|
|
|
|
:param xml: The :class:`~sleekxmpp.xmlstream.stanzabase.ElementBase`
|
|
stanza to compare against.
|
|
"""
|
|
if hasattr(xml, 'xml'):
|
|
xml = xml.xml
|
|
x = ET.Element('x')
|
|
x.append(xml)
|
|
|
|
if not IGNORE_NS:
|
|
# Use builtin, namespace respecting, XPath matcher.
|
|
if x.find(self._criteria) is not None:
|
|
return True
|
|
return False
|
|
else:
|
|
# Remove namespaces from the XPath expression.
|
|
criteria = []
|
|
for ns_block in self._criteria.split('{'):
|
|
criteria.extend(ns_block.split('}')[-1].split('/'))
|
|
|
|
# Walk the XPath expression.
|
|
xml = x
|
|
for tag in criteria:
|
|
if not tag:
|
|
# Skip empty tag name artifacts from the cleanup phase.
|
|
continue
|
|
|
|
children = [c.tag.split('}')[-1] for c in xml]
|
|
try:
|
|
index = children.index(tag)
|
|
except ValueError:
|
|
return False
|
|
xml = list(xml)[index]
|
|
return True
|